2026-11-01 · America/Los_Angeles · 社区动态 · #7

I pushed all 294,912 MLP neurons in Qwen3.5-4B one at a time. Some single neurons turn it into a legal steroid spambot, a bank scammer or a WWE wrestler

The video is real time on my laptop (RTX 5070 Ti, 12 GB). Same question every time. The top line is the plain model, and the answer under it has one neuron pushed. There's no system prompt and no finetuning, the weights don't change. I pushed every MLP neuron of Qwen3.5-4B up and down during generation, one at a time. That's 589,824 runs, and it took about 42 minutes. The neurons that changed the answers most while keeping the text readable then went through a blind test on ten questions they were never picked on, next to random neurons as controls. 130 single neurons passed every check.…

热度 46.0 / 100；排名与评分保留该期记录。